Unsupervised learning for social systems

Many social datasets arrive without clean labels. We often do not know in advance what kinds of students, firms, neighborhoods, conversations, or cultural trajectories should exist in the data.

Unsupervised learning helps make that uncertainty productive. Clustering, embeddings, dimensionality reduction, topic models, and anomaly detection can expose structure that is hard to see with standard summaries. The key is not to treat these methods as automatic discovery engines, but as disciplined ways to generate hypotheses and compare representations.

For CRiSS-LAB, this matters because many of our questions are relational and behavioral: how students organize into groups, how scientific attention decays, how cities constrain movement, and how cultural systems remember. Good unsupervised workflows give us a way to explore those systems without forcing the wrong categories too early.